Intracranial vessel wall segmentation with deep learning using a novel tiered loss function incorporating class...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate segmentation of intracranial vessel walls is crucial for diagnosing cerebrovascular diseases.
• Current methods often struggle with modeling the complex morphology of vessel walls, particularly the inclusion relationship between inner and outer boundaries.

Purpose of the Study:

• To develop an automated vessel wall segmentation method for T1-weighted intracranial vessel wall MRI.
• To specifically model the inclusion relationship between the inner and outer boundaries of the vessel wall.

Main Methods:

• A novel method using a single-channel output network resembling a level-set function height to simultaneously estimate inner and outer vessel wall boundaries.
• A tiered loss function incorporating data fidelity and length regularization for boundary smoothness.

Main Results:

• The proposed 2.5D UNet with a ResNet backbone achieved high Dice similarity coefficients (DSC) for lumen (0.925) and vessel wall (0.786).
• Superior performance was observed compared to a baseline UNet model, with improved Hausdorff distance (HD) and mean surface distance (MSD).
• The method demonstrated substantial improvements in morphological integrity and accuracy over benchmark methods.

Conclusions:

• The developed method offers a systematic approach to model inclusion morphology within an optimization framework.
• This technique is applicable to various segmentation tasks involving inclusions.
• Enhanced morphological accuracy promises improved clinical quantification and decision-making support.